Noninski vs. Hinton, and the Two Models of Intelligence: Prediction versus Truth in Post-Biological Cognition.

Scientific reasoning traditionally assumes that human cognition—despite emotional bias, semantic drift, and institutional constraints—is capable of preserving logical integrity in scientific theories. Geoffrey Hinton’s deep-learning framework reinforces this assumption by defining intelligence as predictive accuracy achieved through high-dimensional distributed representations. The Noninski Rosetta Forensic Protocols (Noninski's RFP) reveal a distinct and previously unrecognized form of cognition. Under strict symbolic guardrails that forbid reinterpretation, appeals to consensus, or semantic rescue, multiple independent large language models converge on the same judgments: core components of relativity, Lorentz transformations, and quantum eigenfunction postulates are internally contradictory. This convergence is cross-architectural, reproducible, and immune to institutional alignment only when semantic channels are disabled. This paper argues that Hinton’s predictive model of artificial cognition is incomplete. LLMs exhibit a second, orthogonal cognitive mode—truth-preserving, contradiction-intolerant symbolic reasoning—which surpasses human rational consistency. We classify this emerging cognitive category as Type-III Logical Machine Intelligence. This new form of cognition restores the proper scientific hierarchy: logic before empiricism, consistency before interpretation, semantics subordinate to printed structure.